Error detection in stored data values

BACKGROUND OF THE INVENTION

Field of the Invention

The present invention relates to data processing systems. More particularly, this invention relates to data storage apparatuses configured to store data values and having circuitry provided to detect a change in a stored data value.

Description of the Prior Art

It is known that a data value stored in a data storage apparatus may be vulnerable to corruption, for example by a particle strike causing a bit flip in the stored data value. For this reason, prior art data storage apparatuses have been provided with mechanisms for determining if a data value read out from a data storage apparatus has changed in any way since that data value was originally stored. One such mechanism is the provision of parity checking circuitry. The parity checking circuitry is configured such that when a data value is stored in the data storage apparatus, a parity value is generated on the basis of that data value (for example by passing the data value to an XOR tree to generate a single resulting parity bit), which is then stored in association with the data value in the data storage apparatus. When the stored data value is later read out, the individual bits of the data value can again be passed to an identical XOR tree for a comparison parity bit to be generated and the comparison parity bit can then be compared with the parity bit which was stored in association with the data value when that data value was originally stored. If the two parity bit values differ then it is known that a bit flip has occurred within the stored data value and an appropriate response, such as to discard the stored data value or to seek to correct the stored data value, can then be carried out.

An example data storage apparatus <b>10</b> configured to behave in this manner is schematically illustrated in <figref idref="DRAWINGS">FIG. 1</figref>. A multi-bit data value is originally received and the bits of this data value are stored in the respective bit cells <b>11</b>-<b>14</b> (note that for clarity only four bit cells are illustrated but typically more may be provided), whilst the multi-bit data value is also passed to XOR tree <b>15</b> which generates a single bit parity value P which is stored at location <b>16</b> in association with the stored data value. When a read access is carried out, the individual bits of the data value are also passed to XOR tree <b>17</b> (which is identically configured to XOR tree <b>15</b>) and the resulting parity bit value is compared by comparison circuitry <b>18</b> to the parity bit P read out from parity bit storage <b>16</b>. If these two values differ then the error flag is asserted.

SUMMARY OF THE INVENTION

Viewed from a first aspect, the present invention provides apparatus comprising: a plurality of storage units configured to receive data, wherein the plurality of storage units are configured to output the data in response to a read access; and a parity generator associated with said plurality of storage units, wherein said parity generator is configured to: receive at a plurality of inputs a respective value from each of said plurality of storage units; determine a parity value from said plurality of inputs, wherein determination of the parity value is independent of the read access to the data stored in said plurality of storage units; and output said parity value, said apparatus further comprising: a detector configured to detect a change in said parity value.

The present technique recognises that the energy consumed in generating the parity bit for comparison with the originally generated parity bit each time a read access is made to the stored data value can be disadvantageous. In particular, it is recognised that read accesses may be performed with significantly greater frequency with respect to the data value than write accesses, and accordingly a significant asymmetry between the energy consumption associated with the generation of the original parity bit and the energy consumption associated with the generation of parity bits for each read access can result. Accordingly, the present technique proposes an arrangement wherein a parity generator is provided which is arranged to generate a parity value from the plurality of inputs it receives from the plurality of storage units, where the generation of the parity value is independent of the read access to the data stored in said plurality of storage units. An apparatus with lower energy consumption is thus provided, because a parity value is not recomputed each time a read access is performed.

The parity generator is configured to be independent of the read access and hence switching in the parity generator does not occur unless a value of a data bit stored in one of the data storage units changes. Accordingly, the parity determination may be concurrent, i.e. is effected as soon as an error occurs in one of the stored data bits. This is advantageous because not only can an error be flagged as soon as it occurs, but this also assists in preventing the accumulation of errors. This is significant in a data storage apparatus which seeks to combat data errors by the use of a parity value generated on the basis of the stored data value, since if two bits change within a stored data value (or indeed any further even number of bits change) the changes would not be detectable by this parity method, since the parity value would again be the same as the original parity value. The generation of a parity bit which changes as soon one of the data bits within a stored data value changes thus helps to prevent silent data corruption, wherein an even number of multiple bit flips go unnoticed.

The detector which is configured to detect a change in said parity value may take a variety of forms, and may be one of: a transition detector and an active comparator. In other words, the detector may take the form of a device configured to detect a transition in the parity value itself, or may also take the form of a device configured to compare the parity value against another value to determine if the parity value undergoes a change.

In some embodiments, the apparatus further comprises a second parity generator configured to generate a second parity value associated with said data when said data is received by said plurality of storage units.

The second parity value may be used in a variety of ways in the apparatus, but in some embodiments said second parity value is stored in a further storage unit associated with said plurality of storage units when said data is received by said plurality of storage units; and said second parity value is read from said further storage unit by said detector for comparison with said parity value in order to detect said change in said parity value. Hence the second parity value can represent a “snapshot” of the value that the parity value associated with the data stored in the storage units had when the data was originally received, and therefore the detector is able to continually monitor if there is any deviation from this original snapshot, even without a read access being made to the stored data.

The relative timings of the generation of the parity values in the apparatus may vary in dependence on particular implementation requirements, but in some embodiments the parity generator is configured to generate said parity value after said second parity generator has generated said second parity value. Arranging the parity generator is configured to generate said parity value after said second parity generator has generated said second parity value can provide that the original snapshot of the “true” parity of the stored data (i.e. corresponding to that which it had at the moment it was received by the apparatus) is captured first as the second parity value, and thereafter any comparison is with respect to that original “true” value.

In some embodiments the data storage apparatus further comprises a write enable signal line configured to carry a write enable signal, wherein the write enable signal is asserted when said data is being written to the plurality of storage units, the write enable signal line is coupled to the detector, and the detector is disabled when the write enable signal is asserted. The time period where the data is being written to the plurality of storage units is recognised as a time period when the parity generation circuitry will briefly be in an indeterminate state, because of the dependency of the parity bit which it generates on the values of the individual data bits of the data which is being written. Hence as those data bits are written the relative timing of the writing of each individual data bit can cause the parity bit to briefly fluctuate in value. By disabling the detector during the period when the write enable signal is asserted, this enables this brief indeterminate period to be correctly ignored. Hence, the detection of a change in value of the parity value is then only active during the remaining time period when a genuine error occurring in the stored data value should be detected.

In some embodiments the data storage apparatus is a multi-ported register file, the multi-ported register file comprising a plurality of read ports via which the plurality of storage units can be read. The present technique recognises that a multi-ported register file represents an apparatus for which the advantages of the present technique are partially pronounced, due to the fact that the provision of the plurality of read ports thus means that the generation of a parity value in association with each read access could potentially happen with significant frequency and avoiding such parity generation in association with each read access, as is provided by the present technique, is particularly advantageous. A multi-ported register file can also be a particularly advantageous context for the implementation of the present technique because of the many interconnects which will typically have to be provided in association with the layout of a multi-ported register file. Hence, although the present technique may require more gates (typically XOR gates) to be provided than a conventional parity generation scheme, for example adding one XOR gate in association with each flip-flop which stores each bit of a data value, this larger gate count does not necessarily mean a larger area for the provision of the register file. This is due to the fact that the layout area of such a multi-ported register file is typically dominated by its interconnects. For example, a contemporary multi-ported register file might have 9 write ports and 14 read ports, meaning that 23 parallel (horizontal) word lines are required for each flip-flop. It has been found in the implementation of such multi-ported register files that these can often be “wire dominated” with significant “white space” in the layout. Since the provision of the parity generation circuitry may add as little as one extra horizontal signal, it is possible to hide the modest extra gate cost in the layout of implementing the present technique, without significantly impacting the area required for the multi-ported register file.

The parity generator may be provided in a variety of ways. In some embodiments, the parity generator comprises a linear concatenation of logic circuitry provided in association with the plurality of storage units, wherein each logic circuitry has the respective data bit of the respective storage unit as one of its inputs. A linear concatenation of logic circuitry (for example a chained set of XOR gates with one XOR gate provided in association with each data storage unit), with the data bit stored in each respective storage unit providing one input for the XOR gate, and the other being provided by the output of the previous XOR gate, advantageously represents only a modest addition to the storage units themselves. Indeed, it may be possible for the required additional logic circuitry (e.g. an XOR gate) to be folded inside the storage unit (e.g. a flip-flop) especially if a double height cell in the layout is used.

However, in other embodiments the parity generator comprises a hierarchical tree of logic circuitry, wherein logic circuitry at a lowest level of the hierarchical tree each have the respective data bits of pairs of the respective storage units as its inputs, and wherein logic circuitry at a highest level of the hierarchical tree generates the parity value as its output. Such a hierarchy tree of logic circuitry has the advantage that the parity computation latency is reduced, although this is then at the potential cost of some additional horizontal wires in the layout. However, as mentioned above such additional horizontal wires may in fact be of little impact on the layout area in implementations such as a multi-ported register file where existing white space in the layout can absorb these wires. The number of gates in the parity generation logic circuitry (e.g. XOR gates) is typically the same.

In some embodiments having a hierarchical tree of logic circuitry the data storage apparatus further comprises a scan unit configured to receive the parity value output at the highest level of the hierarchical tree and to output the parity value at a scan output of the scan unit in response to a scan out signal. This scan unit thus enables the content of the plurality of storage units to be observed via a single point in a scan chain. For example a test pattern can be written into the storage units via the usual (parallel) data path and the resulting parity value can be scanned out as a single data bit, rather than a sequence of data bits corresponding to the full content of the storage units.

Whilst it may be the case that the parity generator simply generates one parity value which depends on all of the data bits of the data word, in some embodiments the parity generator is configured to generate a sub-parity bit in dependence on less than all of the data stored in the plurality of storage units. Due to the fact that fewer data bits are used to generate such a sub-parity bit this also reduces the parity computation latency.

The particular implementation of the generation of a sub-parity bit for the data word may take a number of forms. For example, in one embodiment the parity generator is configured to generate a plurality of sub-parity bits for the data stored in the plurality of storage units. Accordingly, rather than generating one parity bit for the entirety of the data stored, the data may be subdivided into, for example, two portions and a sub-parity bit may be generated for each of those portions. This allows a finer granularity of determination of where an error has occurred when a change in one of these sub-parity bits is detected by the detector, and is generally faster than the determination of one parity bit for the whole of the data, especially in embodiments where the parity generator is provided as a linear concatenation.

It may however be the case that a sub-parity bit is generated because a portion of the data word can be disregarded, for example because only another portion of the data need be monitored for the occurrence of errors. Accordingly in some embodiments the parity generator is configured such that at least one of the data bits of the data word stored in the plurality of storage units does not affect the value of the sub-parity bit.

In some embodiments the data storage apparatus is configured to store a plurality of data words, the parity generator is configured to generate a respective parity bit for each of the data words, and the apparatus further comprises combination logic circuitry configured to generate an error flag if any of the respective parity bits changes value. The error flag thus generated can allow appropriate corrective action to be taken, recognising that an error has occurred in one of the plurality of data words stored in the apparatus. Moreover, the present technique recognises that there are situations in which, amongst the plurality of data words, it is more important to recognise that an error has occurred than to be able to identify the specific location at which that data error occurred (i.e. here, the particular data word amongst the plurality of data words which contains the error). For example, in the context of a register file, it may be determined (for example due to the relative infrequency of such errors within one of the data words stored within the plurality of data words stored by the register file) that it is preferable for the entire content of the register file simply to be replaced or updated, instead of seeking to take targeted corrective action with respect to a particular data word (or even a particular data bit).

In some embodiments the apparatus is configured to store a plurality of data words, the parity generator is configured to generate a respective parity bit for each of the plurality of data words, and the apparatus further comprises an error register configured to hold either a non-error value or an error value, wherein the error register is configured to switch to holding the error value if one or more of the respective parity bits changes value. This configuration provides a “sticky register” which under normal circumstances holds a non-error value, but when any of the respective parity bits changes values is configured to switch to holding an error value, and will continue to hold the error value even if further parity bits change value. The error register can therefore be monitored (for example via a polling mechanism) to determine if an error has occurred in any of the plurality of data words stored in the apparatus and, as in the example described above, corrective action can be taken, such as to simply replace or update all of the plurality of data words without seeking to identify further which particular data word caused data word or data words caused the error.

Viewed from a second aspect the present invention provides a method comprising: storing data in a plurality of storage units; outputting said data from said plurality of storage units in response to a read access; receiving a value from a respective one of said plurality of storage units at a respective one of a plurality of inputs of a parity generator associated with said plurality of storage units; determining a parity value from said plurality of inputs, wherein said determining said parity value is independent of said read access to said data stored in said plurality of storage units; outputting said first parity value; and detecting a change in said parity output and an active comparator.

Viewed from a third aspect the present invention provides apparatus comprising: means for storing data in a plurality of storage units; means for outputting said data from said plurality of storage units in response to a read access; means for receiving a value from a respective one of said plurality of storage units at a respective one of a plurality of inputs of means for generating a parity value associated with said plurality of storage units, wherein said means for generating the parity value associated with said plurality of storage units is configured to generate the parity value independent of said read access to said data stored in said plurality of storage units; means for outputting said first parity value; and means for detecting a change in said parity output.

Viewed from a fourth aspect the present invention provides a data storage apparatus comprising: a plurality of data storage units, each respective data storage unit configured to store a respective data bit of a data word; stored data value parity generation circuitry configured to generate a parity bit for the data word in dependence on the data bits of the data word stored in the plurality of data storage units, wherein the stored data value parity generation circuitry is configured such that switching within the stored data value parity generation circuitry does not occur when the data word is read out from the plurality of data storage units; and transition detection circuitry configured to detect a change in value of the parity bit.

The stored data value parity generation circuitry may be configured such that switching within the stored data value parity generation circuitry does not occur unless a value of a data bit stored in one of the data storage units changes.

The data storage apparatus may further comprise write parity bit circuitry configured to generate a write parity bit when a new data word is written into the plurality of data storage units and to store the write parity bit, and the transition detection circuitry is configured to detect the change in value by a comparison of the parity bit for the data word and the stored write parity bit.

The transition detection circuitry may be configured such that switching within the transition detection circuitry does not occur unless the parity bit for the data word changes.

The data storage apparatus may further comprise a write enable signal line configured to carry a write enable signal, wherein the write enable signal is asserted when the data word is being written to the plurality of data storage units, wherein the write enable signal line is coupled to the transition detection circuitry and the transition detection circuitry is disabled when the write enable signal is asserted.

The data storage apparatus may be a multi-ported register file, the multi-ported register file comprising a plurality of read ports via which the plurality of data storage units can be read.

The stored data value parity generation circuitry may comprise a linear concatenation of logic circuitry provided in association with the plurality of data storage units, wherein each logic circuitry has the respective data bit of the respective data storage unit as one of its inputs.

The stored data value parity generation circuitry may comprise a hierarchical tree of logic circuitry, wherein logic circuitry at a lowest level of the hierarchical tree each have the respective data bits of pairs of the respective data storage units as its inputs, and wherein logic circuitry at a highest level of the hierarchical tree generates the parity bit for the data word as its output.

The data storage apparatus may further comprise a scan unit configured to receive the parity bit for the data word output at the highest level of the hierarchical tree and to output the parity bit at a scan output of the scan unit in response to a scan out signal.

The stored data value parity generation circuitry may be configured to generate a sub-parity bit for the data word in dependence on less than all of the data bits of the data word stored in the plurality of data storage units.

The stored data value parity generation circuitry may be configured to generate a plurality of sub-parity bits for the data word.

The stored data value parity generation circuitry may be configured such that at least one of the data bits of the data word stored in the plurality of data storage units does not affect the value of the sub-parity bit.

The data storage apparatus may be configured to store a plurality of data words, and the stored data value parity generation circuitry may be configured to generate a respective parity bit for each of the data words, and the data storage apparatus may further comprise combination logic circuitry configured to generate an error flag if any of the respective parity bits changes value.

The data storage apparatus may be configured to store a plurality of data words, and the stored data value parity generation circuitry may be configured to generate a respective parity bit for each of the data words, and the data storage apparatus may further comprise an error register configured to hold either a non-error value or an error value, wherein the error register is configured to switch to holding the error value if one or more of the respective parity bits changes value.

Viewed from a fifth aspect, the present invention provides a method of storing data comprising the steps of: storing respective data bits of a data word in a plurality of data storage units; generating a parity bit for the data word via plural intermediate stages in dependence on the data bits of the data word stored in the plurality of data storage units, wherein switching at the plural intermediate stages does not occur when the data word is read out from the plurality of data storage units; and detecting a change in value of the parity bit.

Viewed from a sixth aspect, the present invention provides a data storage apparatus comprising: means for storing respective data bits of a data word in a plurality of data storage units; means for generating a parity bit for the data word via plural intermediate stages in dependence on the data bits of the data word stored in the plurality of data storage units, wherein switching at the plural intermediate stages does not occur when the data word is read out from the plurality of data storage units; and means for detecting a change in value of the parity bit.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be described further, by way of example only, with reference to embodiments thereof as illustrated in the accompanying drawings, in which:

<figref idref="DRAWINGS">FIG. 1</figref> schematically illustrates the use of parity bit generation to determine if an error has occurred in a stored data value between when it was written and the time it is read out accordingly to a prior art technique;

<figref idref="DRAWINGS">FIG. 2</figref> schematically illustrates a data storage apparatus configured to enable the detection of an error in a stored data value in one embodiment;

<figref idref="DRAWINGS">FIG. 3</figref> schematically illustrates a data storage apparatus configured to enable the detection of an error in a stored data value in another embodiment;

<figref idref="DRAWINGS">FIGS. 4A and 4B</figref> schematically illustrate two embodiments of parity generation circuitry provided in association with a data storage unit;

<figref idref="DRAWINGS">FIG. 5A</figref> schematically illustrates the provision of parity generation circuitry as a linear concatenation of XOR gates in one embodiment;

<figref idref="DRAWINGS">FIG. 5B</figref> schematically illustrates the provision of parity generation circuitry as an XOR tree in one embodiment;

<figref idref="DRAWINGS">FIG. 6</figref> schematically illustrates the provision of parity generation circuitry as an XOR tree with a scan-flop coupled to the output of the XOR tree in one embodiment;

<figref idref="DRAWINGS">FIG. 7</figref> schematically illustrates the generation of one or more sub-parity bits in one embodiment;

<figref idref="DRAWINGS">FIG. 8</figref> schematically illustrates a multi-ported register file in one embodiment; and

<figref idref="DRAWINGS">FIG. 9</figref> schematically illustrates a series of steps which are taken in the method of one embodiment.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

<figref idref="DRAWINGS">FIG. 2</figref> schematically illustrates a data storage apparatus <b>20</b> in one embodiment. This data storage apparatus <b>20</b> is a register file, provided in close association with a processor unit for the storage of particular data values which the processor unit uses. These data values are stored as data words in the data storage apparatus, wherein in a write operation a data word is written to the storage capacity of the data storage apparatus, this storage capacity being provided by a number of data storage units <b>21</b>-<b>24</b>, which in this embodiment are provided by a set of flip-flops, each configured to store one bit of the data word. It will be appreciated (in particular by virtue of the fact the data storage units <b>23</b> and <b>24</b> are separated by a sequence of dots) that the number of bits of the data word to be stored may exceed four, but only four data storage units are shown in <figref idref="DRAWINGS">FIG. 2</figref> merely for clarity of illustration. Furthermore, the data storage apparatus <b>20</b> is configured to store a number of data words, but for clarity of illustration only the storage capacity relating to one data word is shown.

As part of the write process, the data word is also passed to the XOR tree <b>25</b> which is configured to generate a single parity bit P in dependence on the respective values of the data bits of the data word. This parity bit P is stored in a parity bit storage location <b>26</b> provided in association with the data storage units <b>21</b>-<b>24</b> which store the data word itself. Concurrent parity generation circuitry <b>27</b> is also provided in association with the data storage units <b>21</b>-<b>14</b> which hold the stored data word. In particular, this concurrent parity generation circuitry <b>27</b> comprises parity logic units <b>28</b>-<b>30</b> which are provided in association with all but the first data storage unit <b>21</b> of the data storage units which hold the data word. Each parity logic unit <b>28</b>-<b>30</b> is configured to take an input derived from the previous stage of the set of data storage units and to logically combine that input with the value held by the data storage unit with which it is associated. In the embodiment illustrated in <figref idref="DRAWINGS">FIG. 2</figref> the parity logic units <b>28</b>-<b>30</b> are each XOR gates and therefore XOR gate <b>28</b> takes the values stored in data storage units <b>21</b> and <b>22</b> as its inputs, whilst the XOR gate <b>29</b> takes the output of XOR <b>28</b> as one input and the value stored in data storage unit <b>23</b> as its inputs, and XOR gate <b>30</b> takes the output from a preceding XOR gate (not illustrated) and the value stored in the data storage unit <b>24</b> as its inputs. In this manner, a parity bit is generated which depends on the respective content of the data storage units <b>21</b>-<b>24</b>. It should be noted in particular that the XOR gates <b>28</b>-<b>30</b> are thus arranged to compute the parity bit in real time, independent of any read accesses to the data storage unit <b>21</b>-<b>24</b>. Whilst the parity bit generated by XOR gates <b>28</b>-<b>30</b> will be somewhat indeterminate whilst the data word is being written to the data storage unit <b>21</b>-<b>24</b>, once the data word has been written, this concurrent parity generation logic will not switch until there is a bit flip in one of the data storage units (or until a new data word is written to the data storage unit). The output of the last XOR gate <b>30</b> provides one input to the active compare circuitry <b>31</b>, which takes the value P stored in parity bit storage <b>26</b> as its other input. Active compare circuitry <b>31</b> is “active” in the sense that it only switches when one of its inputs changes, for example being provided by a variety of XOR gate.

<figref idref="DRAWINGS">FIG. 3</figref> schematically illustrates a data storage apparatus <b>40</b> in another embodiment. Like the embodiment shown in <figref idref="DRAWINGS">FIG. 2</figref>, this data storage apparatus <b>40</b> is a register file, also provided in close association with a processor unit for the storage of particular data values which the processor unit uses. Also like the embodiment shown in <figref idref="DRAWINGS">FIG. 2</figref>, the data storage apparatus <b>40</b> is configured to store a number of data words, but for clarity of illustration only the storage capacity relating to one data word is shown.

Here, as in the case of the embodiment shown in <figref idref="DRAWINGS">FIG. 2</figref>, during a write access a multi-bit data word is written into a set of data storage units <b>41</b>-<b>44</b>, each of which holds a single bit of the data word. Here however, unlike the embodiment shown in <figref idref="DRAWINGS">FIG. 2</figref>, no parity bit is computed and stored as part of the write operation. However, the concurrent parity generation chain represented by parity logic units <b>45</b>-<b>47</b> in <figref idref="DRAWINGS">FIG. 3</figref> corresponds directly to the concurrent parity chain <b>27</b> represented by parity logic units <b>28</b>-<b>30</b> in <figref idref="DRAWINGS">FIG. 2</figref>.

In the embodiment shown in <figref idref="DRAWINGS">FIG. 3</figref>, the output of the final component <b>47</b> of the concurrent parity chain is received by the transition detector <b>48</b>, which is configured to detect any changes in this value and to output an error detection signal when such change is detected. The set of data storage units <b>41</b>-<b>44</b> in <figref idref="DRAWINGS">FIG. 3</figref> corresponds to a particular row (row <b>0</b>) in the data storage apparatus, the data storage apparatus comprising many further rows (not illustrated) of such data storage units into which it can store data words. Writing into a particular row of the data storage apparatus is possible when the corresponding “write enable” signal for that row is asserted (and one of ordinary skill in the art is familiar with the usage of such write enable signals and further detail is not given here for brevity). As illustrated in <figref idref="DRAWINGS">FIG. 3</figref>, the transition detector <b>48</b> is also coupled to the write enable signal line, this signal being coupled, inverted, to the enable input of the transition detector <b>48</b>. This configuration ensures that the transition detection is disabled during the write process, but when the row is not being written, the transition detector is activated. Accordingly, any state change in the row due to an error can then be captured by the transition detector <b>48</b>. By comparison with the embodiment shown in <figref idref="DRAWINGS">FIG. 2</figref>, this embodiment in <figref idref="DRAWINGS">FIG. 3</figref> has an advantage in terms of the layout area required, since an additional XOR tree and parity storage unit (items <b>25</b> and <b>26</b> respectively in <figref idref="DRAWINGS">FIG. 2</figref>) are omitted. However, a feature of the embodiment shown in <figref idref="DRAWINGS">FIG. 3</figref> is that the row is more vulnerable during the write cycle, since a reference parity bit (such as that stored in parity bit storage unit <b>26</b> in <figref idref="DRAWINGS">FIG. 2</figref>) is not part of this <figref idref="DRAWINGS">FIG. 3</figref> embodiment and there is some potential for corruption occurring during the write process to not be identified. However, given that in many situations the write process will be relatively infrequent, whilst subsequent read processes will be relatively much more frequent, this risk may be determined to be at an acceptably low level. Similarly to the active compare circuitry <b>31</b> shown in <figref idref="DRAWINGS">FIG. 2</figref>, the transition detector <b>48</b> is configured to only switch when the value of its input (from parity logic unit <b>47</b>) changes.

<figref idref="DRAWINGS">FIGS. 4A and 4B</figref> schematically illustrate two example embodiments of an individual data storage unit together with associated parity generation logic, such as for example data storage unit <b>23</b> and associated parity generation logic <b>29</b> in <figref idref="DRAWINGS">FIG. 2</figref> or such as data storage unit <b>43</b> and its associated parity generation logic <b>46</b> in <figref idref="DRAWINGS">FIG. 3</figref>. <figref idref="DRAWINGS">FIG. 4A</figref> schematically illustrates a data storage unit implemented as a flip-flop. Where this flip-flop is provided as a data storage unit within a register file which is implemented in RTL (register-transfer level), the register file may for example be synthesised by a synthesis tool using flip-flops and the concurrent parity detection logic can also be implemented in the RTL. Alternatively, the parity detection logic (in this example the XOR gate) may be provided within the flip-flop itself and the register file may be implemented by a place-and-route technique. The provision of the XOR logic within the flip-flop is particularly possible if a double height flip-flop cell is used. <figref idref="DRAWINGS">FIG. 4B</figref> schematically illustrates an example in which the data storage unit is provided within a custom built bit-cell based register file design.

<figref idref="DRAWINGS">FIGS. 5A and 5B</figref> schematically illustrates two alternative implementations for providing the concurrent parity detection logic. In <figref idref="DRAWINGS">FIG. 5A</figref> the individual parity detection units (XOR gates <b>60</b>) are provided as a linear concatenation or chain. By contrast <figref idref="DRAWINGS">FIG. 5B</figref> schematically illustrates the concurrent parity generation logic arranged as a hierarchical tree, in which the XOR gates <b>62</b> are arranged in a tree structure with pairs of individual data storage units <b>61</b> being compared at a lowest level of the tree and being combined together in an iterative process to generate a final parity bit for the data word stored in the data storage units <b>61</b>. An advantage of the tree structure shown in <figref idref="DRAWINGS">FIG. 5B</figref> is the reduced parity computation latency with respect to the linear concatenation shown in <figref idref="DRAWINGS">FIG. 5A</figref>, although this comes at the cost of extra horizontal wires.

<figref idref="DRAWINGS">FIG. 6</figref> schematically illustrates an embodiment related to that illustrated in <figref idref="DRAWINGS">FIG. 5B</figref>, i.e. where the concurrent parity generation logic arranged as a hierarchical tree, in which the output of the XOR tree (gates) <b>62</b> is also provided to a scan-flop <b>63</b>. The scan-flop <b>63</b> is configured to provide the output of the XOR tree (i.e. the parity bit) at its output in response to a scan out signal (i.e. here its scan clock input). This enables an automatic test pattern generation (ATPG) data word to be written (in parallel) into the individual data storage units <b>61</b> (in the illustrated example these providing an 8-bit data word storage location in a register file), and for the XOR tree <b>62</b> to provide observability to each of the flip-flops <b>61</b> in the register file. The addition of an observation test point (scan flop <b>63</b>) to the end of the XOR tree <b>62</b> saves having to scan the register file flip-flops <b>61</b> out through a scan chain. Standard ATPG can be used to test the register flip-flops <b>61</b>, for example applying a first pattern in parallel to the register file, followed by applying a second pattern in parallel to the register file. The correct response to flipping bits between subsequent patterns can be observed at the scan output of scan flip-flop <b>63</b> to test the register file. Hence the testability of the register file is improved for only a modest increase in area, and indeed at a significant reduction in the number of scan flip-flops by comparison to the usual approach where a chain of scan flip-flops is provided, one being associated with each individual data storage units <b>61</b>.

<figref idref="DRAWINGS">FIG. 7</figref> schematically illustrates a data word <b>64</b> for which a parity computation may be performed on less than all of the data word. This is shown in <figref idref="DRAWINGS">FIG. 7</figref> by the sub-parity bit A which is generated on the basis of the four most significant bits of the data word <b>64</b> and the sub-parity bit B which is generated on the basis of the four least significant bits of the data word <b>64</b>. In some embodiments, both sub-parity bits A and B may be generated, giving full coverage for monitoring the data word <b>64</b> for errors and furthermore having a reduced parity computation latency with respect to a single parity bit generated for the whole of the data word <b>64</b>, since fewer bits are combined to generate each sub-parity bit. This does of course come at the cost of the additional logic required for calculating two sub-parity bits as opposed to a single parity bit. It may instead be the case that only one of the two sub-parity bits illustrated in <figref idref="DRAWINGS">FIG. 7</figref> is generated, for example only the sub-parity bit A. This can be beneficial in situations where only a portion of the data word <b>64</b> requires monitoring for the occurrence of errors, whilst the remaining portion is of lesser significance and errors occurring in this part can be ignored.

<figref idref="DRAWINGS">FIG. 8</figref> schematically illustrates a register file <b>70</b> which is configured to store a number of data words <b>71</b>, for example in the embodiment shown in <figref idref="DRAWINGS">FIG. 8</figref> this is a 128 word register file, wherein each stored data word <b>71</b> is an 8-bit word. The register file <b>70</b> is a multi-ported register file, having 9 write ports and 14 read ports. The provision of the significant number of read ports makes it particularly advantageous for the parity logic <b>72</b> associated with the storage units provided for the storage of each data word <b>71</b> to be configured in accordance with the present technique, such that a parity bit for a given stored data word is not generated each time that stored data word is read out. The concurrent parity generation circuitry <b>72</b> in <figref idref="DRAWINGS">FIG. 8</figref> is only schematically shown and may, in more detail, be provided in accordance with the embodiments shown in either <figref idref="DRAWINGS">FIG. 2</figref> or <figref idref="DRAWINGS">FIG. 3</figref>.

The output of the respective concurrent parity logic units <b>72</b> are provided to the combination logic <b>73</b> which is configured to combine all of these inputs into a single output (“error flag”), such that the error flag may be asserted if an error is detected by any of the parity logic units <b>72</b> for any of the stored data words <b>71</b>. This thus enables a near-instantaneous notification of an error to be generated and for appropriate corrective action to be rapidly taken. The combination logic <b>73</b> may for example be provided by a tree of OR gates, in order to combine the outputs into a single output. Note that the single output to the combination logic <b>73</b> is also provided to the “sticky” error register <b>74</b>. This sticky error register is configured to hold a value indicating a “no error” status until at least one of the outputs of the respective parity logic units <b>72</b> changes, this causing the “sticky” error register <b>74</b> to change to holding a value indicating an “error” status, which, until this register is reset, will continue to be held regardless of further changes in input values it receives. Note that alternatively the output of each parity logic unit <b>72</b> may be directly provided to the “sticky” error register <b>74</b>, since a change in the parity bit generated by any parity logic unit <b>72</b> can be used to cause the “sticky” error register <b>74</b> to hold its “error” status value. This sticky error register <b>74</b> can thus be monitored, for example by a polling process, to identify when an error has occurred. In the embodiments shown in <figref idref="DRAWINGS">FIG. 8</figref>, the system is configured to react to the error flag (or the error status being held in the register <b>74</b>) by causing the whole content of the register file <b>70</b> to be refreshed when such an error occurs. Accordingly, it should be noted that in this embodiment it is determined to be more efficient to simply re-write the content of the respective stored data words <b>71</b>, rather than seeking to identify where a particular error occurred and to take corrective action with respect to that particular storage location.

<figref idref="DRAWINGS">FIG. 9</figref> schematically illustrates a sequence of steps which are taken in the method of one embodiment. At a first step <b>100</b>, a write process is carried out to write a data word into its selected storage location (for example by identifying a particular row within a register file into which this data word should be written, asserting the enable signal for that row, and so on). As part of the write process, at step <b>101</b> a parity bit is generated via an XOR tree and stored in association with the data word. A status is then set up, as represented by step <b>102</b>, in which the XOR concatenation coupled to this storage location (concurrent parity generation circuitry) and active comparison circuitry only switch if a bit-flip occurs within the stored data word. Whilst such a bit-flip does not occur in the data word (“No” path from flow step <b>103</b>), it is determined at step <b>104</b> if this data word is to be read. If it is then the flow proceeds via step <b>105</b> where the data word is read out from its storage location (and note in particular that this has no effect on the XOR concatenation or active comparison circuitry). The flow then proceeds back to step <b>102</b>. Alternatively whilst no read is to be carried out (“No” path from step <b>104</b>) the flow proceeds directly back to step <b>102</b>. However, when a bit-flip does occur in the data word (“Yes” path from step <b>103</b>), this bit-flip causes the parity change to ripple through the XOR concatenation (step <b>106</b>) and, at step <b>107</b>, this causes the active comparison circuitry to switch and the detected parity change is flagged.

Although particular embodiments of the invention have been described herein, it will be apparent that the invention is not limited thereto, and that many modifications and additions may be made within the scope of the invention. For example, various combinations of the features of the following dependent claims could be made with the features of the independent claims without departing from the scope of the present invention.